Mapping for Public Health: Initial Plan for Using Satellite Imagery for Micronutrient Deficiency Prediction Elizabeth Bondi Andrew Perrault Fei Fang [email protected] [email protected] [email protected] Harvard University Harvard University Carnegie Mellon University Benjamin L. Rice Christopher D. Golden Milind Tambe [email protected] [email protected] [email protected] Princeton University Harvard University Harvard University ABSTRACT physical manifestations. Wasting is only one type of malnutrition The lack of micronutrients is a major threat to the health and that presents itself in physical manifestations. Therefore, regions development of populations, and it is challenging to detect such with micronutrient deficiency are largely unknown to public health deficiency at large scale with low cost. In this work, we planto organizations until direct measurements are made, such as blood use data from a study on micronutrient deficiency in Madagascar, draws. However, these blood draws and questionnaires are costly which include blood draw results and corresponding questionnaires, and time-consuming, and furthermore, quantifying micronutrient along with satellite imagery, to determine whether there are certain levels in a sample requires specialized laboratory equipment that cues visible in satellite imagery that could more easily and quickly is not widely available. We envision a new approach to detect mi- suggest areas where people may be susceptible to micronutrient cronutrient deficiency at large scale with low cost. deficiency. We propose an approach that will (i) determine impor- We hypothesize that the availability of foods with sufficient mi- tant predictors of micronutrient deficiency from blood draws and cronutrients in a community depends in part on environmental corresponding questionnaire data, such as type of food consumed, factors, such as the state of forests or agriculture nearby. Madagas- (ii) automatically detect related areas in satellite imagery, such as car is a unique area for the purposes of studying this hypothesis, forest regions where sources of important food may be found, and as it has many different ecological regions within it [15]. A recent (iii) use these to predict regions of micronutrient deficiency. As the cross-sectional study examined micronutrient deficiency in sev- prediction of micronutrient deficiency in satellite imagery will be eral of these regions, in addition to anemia, malaria, and detailed done using meaningful predictors, and we anticipate using an in- questionnaire data [9], which we describe further in Section 3. We herently interpretable model for prediction based on these objects, propose an approach that will use satellite imagery to identify im- such as logistic regression, we aim to create a model that will be portant environmental factors that may help predict regions of intuitive to those in the public health community.1 micronutrient deficiency. With such an approach, interventions such as supplementation, CCS CONCEPTS fortification, biofortification, and infectious disease management, could be targeted based on more up-to-date information [2]. In • Computing methodologies ! Computer vision problems; brief, supplementation, e.g. vitamins, is costly and therefore more Machine learning; • Applied computing ! Life and medical targeted at an individual level, while fortification, e.g., adding iodine sciences. to salt, is inexpensive and could be applied to regions, though it KEYWORDS requires population-wide deficiencies and an inability to reach toxic limits with too much of the micronutrient. Biofortification interpretability, remote sensing, micronutrient deficiency is similar to fortification in that it seeks to add nutrition to foods, ACM Reference Format: but in this case adds nutrients during growth, e.g., adding iron Elizabeth Bondi, Andrew Perrault, Fei Fang, Benjamin L. Rice, Christopher to crops during growth. Finally, infectious disease management is D. Golden, and Milind Tambe. 2020. Mapping for Public Health: Initial Plan useful to curb micronutrient deficiency, as infectious diseases often for Using Satellite Imagery for Micronutrient Deficiency Prediction . In deplete micronutrients. Specific interventions would depend on the KDD 2020 Workshop on Humanitarian Mapping. disease(s). Throughout the rest of this proposal, we will provide a brief literature review for different aspects of this hypothesis, 1 INTRODUCTION including literature in artificial intelligence and in public health, According to the World Health Organization, the lack of micronu- highlight the proposed methodology, and conclude by summarizing trients “...represents a major threat to the health and development our vision. of populations the world over, particularly children and pregnant women in low-income countries.” [1]. However, micronutrient de- 2 RELATED WORK ficiency is sometimes called “hidden hunger”, as the effects often First, we reviewed several potential uses of AI in the context of become visible only when the deficiency is already severe [29]. In nutrition, including AI and malnutrition and AI and micronutrient other words, the majority of the types of malnutrition do not have deficiency. There is work on predicting anemia and/or hemoglobin 1KDD 2020 Workshop on Humanitarian Mapping. Authors retain the copyright. levels [6, 8, 24, 32], as well as predicting mutations that lead to Bondi, et al. Vitamin B12 3 PROPOSED METHODOLOGY Vitamin A Iron We propose to (i) use the questionnaire data, which specifically con- Question 1 sist of responses of 6292 individuals from 1125 households within Question 2 Question 3 24 communities in four of the ecological regions of Madagascar, ... to determine what predicts micronutrient deficiency in the 2881 individuals who also have micronutrient data, then (ii) use these Questionnaire Micronutrients Data predictors to determine relevant satellite data products to include and/or derive, such as forest cover [28], climate [14], agriculture High Resolution Population (e.g., determined via computer vision techniques), and/or roads NIR Weather RGB Forest [17], and finally (iii) use these products to predict micronutrient deficiency. We plan to use logistic regression for (i) and (iii) tofos- ter interpretability, and deep learning to create/interpret relevant satellite data products in (ii). This is illustrated in Fig. 1. To compute intermediate satellite image products in (ii), we will Satellite Data Satellite Data likely use segmentation techniques, e.g., U-Net [23], to determine (Products) (Raw) exact regions of forest, road, agriculture, etc. within a satellite image. We plan to use information from all respondents (about 6000) to Figure 1: (i) Starting in the first row, we will use question- provide labels for these intermediate products, and to augment naire data to determine what factors can predict micronu- our training data through the use of additional datasets such as trient deficiency. (ii) We will then use these predictors tode- the Demographic and Health Survey/Malaria Indicator Survey in termine the satellite data products that are necessary (dotted Madagascar [30]. The most recent survey includes anemia results arrow). (iii) We will use these products to predict micronu- and information collected regarding food and lifestyle, such as roof trient deficiency. material, which may be used as additional labels. Training and testing data would therefore be done on regions within Madagascar, with hopes of creating a more widely generalizable process for use in other places. To predict micronutrient deficiency in (iii), we will disease related to micronutrients [12]. However, these works do not train using 2000 inputs derived from the intermediate products in consider the prediction of micronutrient deficiency at a regional (ii) and their corresponding micronutrient result, and test on the level using satellite imagery. For malnutrition, there is a great deal remaining 881 micronutrient results. of work on detecting malnutrition in children via smartphone apps We hypothesize that several environmental factors may be par- and computer vision, as well as questionnaires (e.g., [11, 25, 26]). ticularly important, including the presence and health of forests and Similarly, these do not use satellite data, nor do they focus at a agriculture. We also believe that micronutrient deficiency and its regional level. potential environmental influences could be confounded by poverty, There is another body of work on food security in the context so plan to include poverty measures (e.g., from [10]) to control for of agriculture and food rescues, as well as foodborne illnesses [27]. these effects. Furthermore, we hypothesize that easy access to larger There is also a great deal of work on AI and agriculture, for example food markets reflected by proximity to roads would lead to less in aiding smallholder farmers [16], finding crop disease [20], and micronutrient deficiency. If this proved to be true in our analysis, developing systems to improve agricultural output [4, 7]. While for example, it may indicate that there is increased participation in these works may apply at a wider level than those mentioned pre- markets or healthcare along roads, and interventions could focus viously, they still do not make use of satellite imagery, nor do they on widening access to these necessities nearby, for example via specifically pertain to micronutrient deficiency. supplementation. As drawing this conclusion will rely heavily on One main